feat: add lance_dataset_add_columns for SQL / all-null / stream column addition

[LuciferYang]

Summary

Last of three PRs against #41. Exposes upstream's Dataset::add_columns through the three NewColumnTransform cases that translate cleanly across the C ABI:

• SQL expressions — derive new columns from SQL over existing columns.
• All-null columns — add nullable columns from an Arrow schema. On the modern format this is metadata-only; the legacy format can't represent it that way and returns LANCE_ERR_NOT_SUPPORTED.
• Stream — splice in precomputed column data from an Arrow C stream, aligned positionally to the dataset's existing rows.

Upstream's fourth variant, BatchUDF, is left out on purpose: it carries a Rust closure that can't cross the C ABI, and the stream variant already covers the same "bring your own computed data" use case.

Each call mutates the dataset in place under an exclusive write lock; scanners already in flight keep their pre-add view via the same Arc clone-on-write as the _drop_columns (#42) / _alter_columns (#44) siblings.

Three focused entry points rather than one mode-tagged function, because the inputs are genuinely different shapes (name/expression pairs vs. a schema pointer vs. a stream). Upstream's read_columns parameter isn't exposed — it only feeds BatchUDF; for these three variants upstream ignores it. batch_size is forwarded where it does something (SQL scan, stream alignment) and omitted from the metadata-only all-null path.

Surface

typedef struct LanceSqlColumn {
    const char* name;
    const char* expression;
} LanceSqlColumn;

int32_t lance_dataset_add_columns_sql(
    LanceDataset* dataset, const LanceSqlColumn* columns,
    size_t num_columns, uint64_t batch_size);

int32_t lance_dataset_add_columns_nulls(
    LanceDataset* dataset, const struct ArrowSchema* schema);

int32_t lance_dataset_add_columns_stream(
    LanceDataset* dataset, struct ArrowArrayStream* stream,
    uint64_t batch_size);

batch_size uses 0 for the upstream default and is range-checked to u32. Two error-code details are worth calling out, both matching existing behavior: a SQL expression referencing a non-existent column surfaces as LANCE_ERR_INTERNAL (an upstream schema error, the same path as lance_dataset_delete), whereas a syntax error is LANCE_ERR_INVALID_ARGUMENT.

Because these are unsafe extern "C" entry points under panic = "abort", the stream variant pre-validates the mandatory CADI callbacks before handing the stream to arrow-rs (which would otherwise abort on a NULL get_schema / get_next), and the all-null variant rejects an uninitialised or non-UTF-8 top-level schema format before arrow-rs's assert!/expect can fire. The stream is consumed (released) on every non-NULL return path.

Tests

Rust integration tests cover all three variants end to end: computed values (single and multi-column SQL, constant expressions, honored batch_size), all-null backfill, and multi-fragment stream alignment — plus the full rejection surface (NULL/empty/non-UTF-8 inputs, name collisions, row-count mismatch, invalid batch_size, released/missing-callback streams, non-nullable all-null fields, and the legacy-format NOT_SUPPORTED path). Stream-consumption is proven with a drop counter rather than a vacuous release-slot check. C and C++ smoke tests exercise the SQL happy path and each variant's argument rejections across the ABI.

[LuciferYang]

Thank you @jja725